Using a .safetensors T5

[rkfg]

Just a quick note, most probably people already have a T5 in safetensors format, cloning the PixArt repo (41 GB on disk) is quite an overkill. So it might be better to use the regular CLIP loader like this (top from the example workflow here, bottom is what I propose):

The result is slightly different, must be some rounding errors, otherwise it's 99% identical. The other CLIP types produce heavy bright artifacts so stick to the SD3 type.

[kvochko]

@rkfg thank you! Please note that in latest ComfyUI you should use ltxv for model type.

[rkfg]

For me this lxtv type produces noticeably worse results. SD3 or the original loader are much better. Could be an implementation quirk?